Open Stack

On Fri, 2013-11-15 at 11:28 -0500, Russell Bryant wrote:
> Greetings,
> 
> We've talked a lot about requirements for new compute drivers [1].  I
> think the same sort of standards shold be applied for a new third-party
> API, such as the GCE API [2].
> 
> Before we can consider taking on a new API, it should have full test
> suite coverage.  Ideally this would be extensions to tempest.  It should
> also be set up to run on every Nova patch via CI.
> 
> Beyond that point, now is a good time to re-consider how we want to
> support new third-party APIs.  Just because EC2 is in the tree doesn't
> mean that has to be how we support them going forward.  Should new APIs
> go into their own repositories?
> 
> I used to be against this idea.  However, as Nova's has grown, the
> importance of finding the right spots to split is even higher.  My
> objection was primarily based on assuming we'd have to make the Python
> APIs stable.  I still do not think we should make them stable, but I
> don't think that's a huge issue, since it should be mitigated by running
> CI so the API maintainers quickly get notified when updates are necessary.
> 
> Taking on a whole new API seems like an even bigger deal than accepting
> a new compute driver, so it's an important question.
> 
> If we went this route, I would encourage new third-party APIs to build
> themselves up in a stackforge repo.  Once it's far enough along, we
> could then evaluate officially bringing it in as an official sub-project
> of the OpenStack Compute program.

I do think there should be a high bar for new APIs. More than just CI,
but that there is a viable group of contributors around the API who are
involved in OpenStack more generally than just maintaining the API in
question.

I don't at all like the idea of drivers or APIs living in separate repos
and building on unstable Nova APIs. Anything which we accept is a part
of OpenStack should not get randomly made unusable by one contributor
while other contributors constantly have to scramble to catch up. Either
stuff winds up being broken too often or we stifle progress in Nova
because we're afraid to make breaking changes.

I happened to meet Thijs Metsch, the maintainer of OCCI for Nova
yesterday:

  https://github.com/tmetsch/occi-os
  https://wiki.openstack.org/wiki/Occi

and he described how often he had to fix the API implementation to adapt
to changes in Nova's API. My advice was to work towards having the API
be included in Nova (understand that it's a long road and there would be
a bunch of difficult requirements) or take the less technically
attractive option of proxying the OCCI API to the stable OpenStack REST
API.

For someone like Thijs, choosing a middle road where the API
implementation is going to be constantly broken is going to suck for him
and his users.

Mark.